Data Quality & Provenance
Every record returned by every Silicon Analysts tool — REST or MCP — carries a provenance block describing where the data came from, how confident we are in it, and when it was last refreshed. This page is the canonical reference for what those fields mean.
Where our data is least certain
We publish ranges, not false precision. These are the datapoints where our published range is widest — where the real number is hardest to pin down, and where input from people who work with these numbers would move us most. Confidence here is inferred from the width of the published range.
| Datapoint | Published | Range & confidence | Spread | What’s your number? |
|---|---|---|---|---|
| iot-edgegross margin | 45% | 35%–50% | 33% | |
| Intel 16nmwafer price | $4,500 | $3,800–$5,200 | 31% | |
| Samsung 5nm (SF5)wafer price | $13,000 | $11,000–$15,000 | 31% | |
| TSMC 28nmwafer price | $3,000 | $2,550–$3,450 | 30% | |
| TSMC 7nm (N7)wafer price | $9,500 | $8,100–$10,900 | 29% | |
| TSMC 5nm (N5/N4)wafer price | $18,500 | $16,000–$21,300 | 29% | |
| Samsung 3nm (SF3)wafer price | $15,000 | $13,000–$17,000 | 27% | |
| TSMC 3nm (N3)wafer price | $19,500 | $17,000–$22,000 | 26% |
Drop your own number in any row to see how it compares to the crowd — input is anonymous and aggregated (individual values are never shown; the distribution unlocks at 3 contributions). Have a source? Logged-in users can suggest a reviewed correction — how contributions are handled.
How community data works
We invite the people who use these tools to sharpen the numbers — but a number nobody can trust is worse than no number. Here’s exactly how contributions are handled.
Anonymous by default
No account or email needed. We store only the value and which datapoint it’s for, to compute a community range — never tied to your identity, never shown individually. The range also reflects adjustments people make in our tools; you can opt out of being included at any time.
The crowd informs, we decide
Contributions never change our published numbers automatically. Sourced corrections go to a review queue we vet by hand; implausible values are bounded out, and any stray number is absorbed harmlessly into a median — it can’t move what we publish.
Three labels, never blurred
A number is always one of: our published estimate (vetted), the community signal (n contributors, median + confidence), or validated (a sourced correction we reviewed). We never present crowd input as fact.
Showing our confidence — and where the crowd disagrees — isn’t a weakness. It’s how we stay honest about what’s known versus estimated.
Provenance Taxonomy
Two enums classify every record: source_type (how the record was produced) and confidence_tier (how much to trust it). The type contract lives in lib/tools/types.ts.
source_type — how this record was produced
Sourced from a published external report (Morgan Stanley, TrendForce, public earnings, etc.). Direct attribution.
Computed by Silicon Analysts from public inputs via a documented methodology (Monte Carlo cost models, yield equations, etc.).
Pure-function output from user-supplied inputs (e.g. calculate_chip_cost results). No data lookup involved.
Analyst judgment where data is sparse or unobservable.
confidence_tier — qualitative confidence
Multiple independent sources agree; methodology is well-tested.
Single authoritative source or moderate methodology uncertainty.
Sparse data, significant assumptions, or rapidly changing.
0.85 is worse than a reasoned "high" — agents will treat the number as more precise than it is. We will tighten to a numerical confidence once the methodology for computing it is documented.Methodology & Models
The calculators run documented, industry-standard models — published here (and in /llms.txt for AI agents) so anyone consuming a number can assess how it was produced.
Yield models
User-selectable Poisson Y = e−A·D0 and Murphy Y = ((1 − e−A·D0) / (A·D0))2, where A is die area in cm² and D0 is defect density in defects/cm².
Defect-density maturity
D0 scales by production phase (risk / ramp / high-volume). The risk factor derives from tracked per-node early-vs-mature defect densities (e.g. TSMC N3: 0.14 early vs 0.09 mature ≈ 1.56×), with a documented 1.4× fallback when a node lacks an early figure; ramp uses the geometric midpoint (√ratio). A small assembly-yield hit applies in immature phases (risk ×0.97, ramp ×0.99). These phase factors are SA estimates and labeled as such in the tools.
Chiplet economics
Multi-die designs use known-good-die (KGD) costing — default KGD yield 98.5%, packaging cost scaled +15% per die beyond the first, up to 4 dies, with assembly yield compounding per die. Optional 3D stacking compounds per-layer bond yield: Ystack = Ybond(layers−1) (US8466542), 1–8 layers, default 100% (an exact no-op). Bond-yield presets: mature 98% (ECTC 2026); ramp/early-ramp figures are SA estimates — Samsung's HBM4 hybrid-bonding ramp has been publicly reported near ~10%.
Manufacturing-energy adder (opt-in)
kWh per wafer start by node (~200–400 legacy up to ~800–1,500 at 3nm; ASML/imec/McKinsey public figures) × regional industrial electricity rate across 7 fab regions (EIA, Taipower, KEPCO tariffs) × an optional ~1.75× facility-overhead multiplier (SA estimate — facility systems draw 50–57% of total fab electricity). Wafer price already embeds foundry energy, so the adder is a regional attribution / scenario delta, not an additional BOM line. Off by default; labeled SA est. wherever shown.
Explicit gaps — not modeled
Two packaging-economics quantities have no credible public figure and are deliberately not in the model: the hybrid-bonding cost delta vs micro-bump in $/wafer (only tool-throughput proxies exist — noted qualitatively on SoIC-class entries), and thermal-derate-to-cost rules (power-density ceilings stay qualitative). We document gaps rather than invent parameters.
Wafer price consensus
Multi-source corroborated wafer prices are rounded to $250 steps and published as bands (min/avg/max) — deliberate precision honesty. The what-changed feed applies a 1% significance floor on wafer pricing so rounding wobble never reads as a real movement.
Update Cadence by Dataset
Last-updated dates below are dataset-level. Per-record overrides are supported in the API; future migrations will populate them as individual chips, nodes, and packaging types refresh asynchronously. Cadence is a target — historical refresh history will appear on a planned /changelog page.
| Dataset | Last Updated | Target Cadence | Source Types |
|---|---|---|---|
AI Accelerators (chipSpecs) Per-chip cost breakdowns derived from Monte Carlo models against public teardown data and analyst reports. | 2026-04-07 | Monthly | derivedresearch |
Foundry & Wafer Pricing (foundryData) Wafer price ranges, defect density, NRE/mask-set costs, and node maturity status. Synthesized from TrendForce, Morgan Stanley, CSET, public filings. | 2026-06-22 | Quarterly | derivedresearch |
Packaging & HBM Specs (packagingData) Per-tech packaging cost benchmarks (CoWoS-S/L, EMIB, SoIC, FC-BGA, etc.) plus HBM2 → HBM4 cost-per-stack and bandwidth/capacity. | 2026-07-05 | Monthly | derivedresearch |
HBM Market Analysis (hbmData) 9 sub-tables: accelerators, specs, market share, spot prices, leading indicators, qualification feed, revenue forecast, supplier revenue, validation checks. | 2026-04-07 | Monthly | researchderived |
Supply-chain Headlines (marketPulse) Curated supply-chain headlines with trend direction and impact analysis. Per-item dates parsed to ISO 8601 in the API. | — | Weekly | research |
Methodology Notes by Source Type
Research
Records attributed directly to a public external publication. Primary sources include TrendForce quarterly reports, Morgan Stanley semiconductor research, Raymond James analyst notes, CSET, IEDM proceedings, and earnings releases. Each record carries a per-row source string in addition to the structured provenance block.
Derived
Records produced by a Silicon Analysts model from public inputs. Examples: per-accelerator cost breakdowns combine Epoch AI Monte Carlo models with TrendForce and Raymond James inputs; wafer price ranges synthesize multiple foundry-pricing sources with the Murphy yield model. See Semiconductor Cost Guide for the methodology in depth.
Computed
Pure-function output from user-supplied inputs. The calculate_chip_cost tool is the only example today: given die dimensions and process parameters, returns an estimated chip cost. No data lookup is involved at the record level (though input defaults pull from derived wafer-pricing data — that data's freshness sets the ceiling on result freshness via the conservative-pick rule).
Estimated
Analyst judgment where data is sparse or unobservable. Rare today; reserved for fields like packaging cost on early-life tech where no published price exists. Always paired with low or medium.
Example: Provenance in a Tool Response
Every record across all 15 tools carries the same shape. Below is a representative get_accelerator_costs record (truncated).
{
"chip": "NVIDIA B200",
"vendor": "NVIDIA",
"processNode": "TSMC N4P",
"estMfgCostUsd": 8500,
"estSellPriceUsd": 30000,
"costBreakdown": { "logicDieCostUsd": 220, "hbmCostUsd": 4200, ... },
"provenance": {
"last_updated": "2026-04-07T00:00:00.000Z",
"source_type": "derived",
"confidence_tier": "high",
"dataset_version": "chipSpecs-v1.0"
}
}See the Developer API page for full per-tool response shapes and the /api/v1 manifest for the machine-readable contract.
Related
Immutability & validation
The record is written in pen, not pencil. Here is exactly how a number gets published, why it can't be silently changed afterward, and what happens when one turns out to be wrong.
A frozen daily ledger
Every day since 6 June 2026, the day's published values are photocopied into an append-only ledger. Those rows can never be edited or deleted — the database itself rejects the attempt, even from our own admin account. You can replay any past date exactly as it stood via /api/v1/snapshot?date=YYYY-MM-DD. History is never backfilled or rewritten.
Four checks before anything is published
A value must clear every gate to reach the record: (1) database bounds reject impossible values at the source; (2) a pre-freeze validator refuses to freeze an impossible value — the ledger takes a visible one-day gap instead of a permanent wrong number, and an operator alert fires; (3) automated agents only publish a figure when at least two independent public sources agree; (4) any human-approved edit is re-validated against the same bounds at approval time. A value that fails the freeze two days running is auto-reverted to the last frozen valid value — and that revert is itself logged as a public correction.
Corrections are public, never silent
When a published value later proves wrong, it is corrected in the open: an immutable brief is archived and a corrected one re-published; a frozen ledger day is annotated with the affected date range. Every correction is recorded in the public changelog below and served inline by the API for the dates it touched. Sourcing is public-only — no insider data — so every claim is checkable at its source.
Corrections changelog
Published data is immutable — frozen ledger days and event timelines are never edited. When a value turns out to be wrong, the correction is recorded here (and served inline by /api/v1/snapshot, /api/v1/export and /api/v1/changes for the affected dates).